fix(scheduler): correct daemon_status recency, RTW-A provenance tiers, and eval_provider

[bug-ops]

Summary

Three independent zeph-scheduler bugs surfaced by live-testing, grouped into one PR per project triage convention (same subsystem, no cross-dependencies):

• daemon_status() recent_runs was alphabetical, not recent, and last_run was always empty. ScheduledTaskInfo now carries last_run, list_jobs_full selects it, and daemon_status sorts by last_run descending (never-run jobs last) before truncating to recent_n. The CLI printer renders last: alongside next:.
• TaskProvenance::is_external() was dead code. Every RTW-A reentry-defense mechanism gated on != TaskProvenance::Static, collapsing the documented three-tier trust model (Static/UserAdded/External) into a binary one. The injection-pattern-check mechanism now applies unconditionally to External-provenance tasks regardless of the injection_pattern_check config toggle, while UserAdded still respects the toggle — a real strictness differential, without weakening the existing baseline for either tier.
• Scheduled/cron experiment runs ignored eval_provider, judge equaled subject. The interactive /experiment command and --experiment-run CLI flag already resolved a distinct judge via build_eval_provider(); the scheduler path never did, silently defeating the self-judge-bias mitigation in the unattended code path. ExperimentTaskHandler now resolves eval_provider the same way at both scheduler-bootstrap call sites (src/runner.rs, src/acp.rs), falling back to the primary provider only when unset.

Incidental discovery, fixed in scope: src/scheduler.rs's entire mod tests (including pre-existing tests, not just the new ones added here) was gated behind a vestigial #[cfg(feature = "testing")] that CI's actual build-tests archive step (.github/workflows/ci.yml:211) never enables — the module was invisible to real CI regardless of correctness. Regated to #[cfg(all(test, feature = "scheduler"))]; this is what makes the new #5947 regression test actually protect CI rather than being cosmetic. Zero runtime behavior change, test-code only.

Known residual limitation (not fixed here, follow-up filed): the TaskProvenance::External differential added for #5950 is only meaningful against an attacker who honestly self-labels — provenance is trusted-as-read from the DB with no re-verification at hydration, so a direct-SQL external actor can spoof provenance="user_added"/"static" to dodge it. Not a regression (pre-fix both non-Static tiers were equally bypassable); tracked as #6114.

Follow-up issues filed

• scheduler: TaskProvenance is trusted-as-read from DB, allowing External tasks to self-label as UserAdded/Static #6114 — force External provenance on all DB-hydrated rows instead of trusting the stored column (closes the spoofing gap noted above)
• zeph-scheduler: daemon_status fetches all jobs then sorts in Rust instead of pushing ORDER BY/LIMIT into SQL #6115 — push daemon_status's ORDER BY last_run DESC LIMIT into SQL instead of fetching all jobs and sorting in Rust (current approach avoids sqlite/postgres NULLS LAST portability risk but doesn't scale query-side)

Test plan

• cargo +nightly fmt --check — clean
• cargo clippy --profile ci --workspace --all-targets --features "desktop,ide,server,chat,pdf,scheduler,testing" -- -D warnings — clean
• cargo nextest run --config-file .github/nextest.toml --workspace --features "desktop,ide,server,chat,pdf,scheduler" --lib --bins — 12983 passed, 0 failed, 35 skipped
• RUSTFLAGS="-D warnings" RUSTDOCFLAGS="--deny rustdoc::broken_intra_doc_links" cargo doc --no-deps --workspace --features "desktop,ide,server,chat,pdf,scheduler" — clean
• cargo test --doc -p zeph-scheduler --features "sqlite,daemon" — 23 passed
• New regression test experiment_task_handler_uses_distinct_eval_provider confirmed to fail on revert (two distinct MockProviders, asserts independent recorder traffic for subject vs. judge)
• New daemon_status recency-ordering test (older/newer/never-run, asserts order and actual values)
• Two new differential tests proving External vs UserAdded RTW-A treatment, plus unweakened baseline regression coverage
• .local/testing/coverage-status.md and .local/testing/playbooks/scheduler-cli.md updated with scenarios for all three fixes (Untested status pending live-session verification in a future CI cycle)

Closes #6096
Closes #5950
Closes #5947